Memory for Text Research Paper

1. Relation Between Processing and Memory for Text

Making inferences and integrating information play important roles in the process of constructing a mental representation during comprehension. In making inferences, readers activate knowledge from their semantic memory that is relevant to, but left implicit in, the text. Integration consists of connecting concepts or propositions to earlier propositions by searching episodic memory, and may be followed by making an inference. The memory representation depends upon these processes. Little, however, is known about the extent to which readers actually perform inference and integration activities during reading. In particular, little attention has been paid to the way in which activation of knowledge in a reader’s memory is regulated. A similar problem arises with regard to the construction and updating of situation models (representations of the situation the text refers to or is about) in episodic memory. To what extent do readers construct and update precise situation models, and what factors influence these processes? This research paper seeks to contribute to a better understanding of the regulation processes involved in making inferences and integrations, in constructing and updating of situation models during comprehension, and, subsequently, influencing memory for text.

2. Completeness of the Representation

One block of factors that are relevant to the regulation processes involved in making inferences and integrations are person-related factors. These factors, such as reading goal and habitual reading style of an individual reader, determine the manner in which inferencing and integrating occur. Readers with a high or strict comprehension criterion infer concepts related to a schematic structure (or script) underlying a story and integrate concepts to a greater extent than do readers with a low or careless comprehension criterion. It seems that if a reader is interested in minimal comprehension the standards for coherence are met relatively easily and little activation of background knowledge takes place. They read, for instance, sentences with script arguments (for example, booking office in the context of a train journey story) in a script-based story faster than sentences with nonscript arguments (e.g., bookstall), regardless of whether the arguments have been mentioned before in that story or not. In contrast, if a reader is interested in attaining a thorough understanding of a text, the standards for coherence are very demanding, reading is slow and involves extensive recruiting of background knowledge or of information from the mental representation that has been constructed so far. These slow or careful readers can be characterized as readers who completely instantiate scripts and also integrate encountered concepts in the episodic memory representation. The outcomes of a recognition task presented afterwards, consisting of judging the implicit script arguments as ‘new’ or ‘old,’ as Van Oostendorp (1991) showed, support this interpretation. Slow or careful readers infer the scriptal arguments in the implicit condition—when the script argument was not mentioned—while fast or careless readers do not. They are, therefore, slower in judging the script argument as new because they have similar or related information represented in episodic memory, making the judgment ‘new’ difficult and slow. These findings indicate how processing influences the memory representation that is being stored.

Regulation also occurs on the basis of textual characteristics. On a sentence level, for instance, it appears that subjects process concepts in semantically less-related sentences more extensively than concepts in semantically high-related sentences (Van Oostendorp 1994). Semantic relatedness was here assessed—among other measures—by means of a rating task in which subjects judge the meaning overlap within each pair of content words in a sentence. Subjects read context sentences such as ‘The cat caught a mouse in the kitchen’ (which contains highly related concepts) as opposed to ‘The cat seized a mole in the field’ (less related). These sentences were also embedded in stories. Immediately after reading such a context sentence, a verification question was presented which referred to a relevant attribute of one concept (has claws for cat). These attributes, particularly low-typical attributes, are verified faster after reading semantically less-related context sentences than after reading highly related sentences. Also the cued recall performance of readers is better for sentences that are semantically less related than for high-related sentences. Semantic high-relatedness may, thus, lead to less activation of knowledge, result in the failure to make inferences and, consequently, to a less elaborate episodic memory trace. In a study by Cairns et al. (1981), subjects read sentences with a predictable or with an unpredictable target word in relation to a preceding context. An example of a predictable word is ‘gum’ in ‘Because she was chewing so loudly in class, Sarah was asked to get rid of her gum promptly.’ It is unpredictable in ‘Because it was annoying the others, Sarah was asked to get rid of her gum promptly.’ The reading time for the second sentence is longer than for the first type of sentence. The recognition and reproduction of unpredictable word sentences is also better than of predictable word sentences. Cairns et al. (1981) assume that more knowledge is activated while processing the former sentence, which leads to prolonged processing and to a more elaborate propositional representation. Semantic relatedness may even induce failures to notice errors in sentences. Van Oostendorp and De Mul (1990) presented subjects with sentences such as ‘Moses took two animals of each kind on the Ark. True or false?’ A majority of subjects answer erroneously ‘true’ although the subjects know the correct name (Noah), as was shown by a later test. Furthermore, sentences with high-related inaccurate names (Moses) lead to more semantic illusions than sentences with low-related inaccurate names (e.g., Adam).

One may conclude that readers continuously monitor the semantic cohesion of a mental representation under construction and regulate further processing on the basis of a comparison of the perceived cohesion to some internal comprehension standard. Myers and O’Brien (1998) take a similar standpoint on the control of processing. During initial processing of a sentence, the perceived cohesion of a propositional representation is primarily dependent on the semantic relatedness between involved concepts (Van Oostendorp 1994, see also Kintsch 1974, p. 214 for the same idea). Often, readers have the competence to making inferences and integrations, but frequently they don’t completely employ this capacity because the initial perceived cohesion is above the standard readers have set.

3. Updating Mental Representations

The same problem with regard to the completeness of the propositional representation can be raised concerning the construction and updating of situation models. Do readers construct and represent in memory detailed situation models under naturalistic conditions, that is, with a more naturalistic text and with a more naturalistic reading task than those often used in laboratory studies? And, also, do they accurately update their model when reading new (correcting) information? Readers often do not form integrated spatial-situation models during comprehension of a naturalistic story, even when they have the opportunity to reread the text, nor do they update their model accurately in episodic memory (Zwaan and Van Oostendorp 1994). The text used in these studies was a part of a detective story. Only when readers are specifically instructed to construct a mental map of the situation described in the story do they form and update spatial situation models. Other studies basically confirm these findings (e.g., Hakala 1999).

Under certain circumstances, a text with too much coherence can also be detrimental to constructing an adequate situation model. A text that is fully explicit and coherent at both the local and global level may result in impaired comprehension at a situation-model level—measured by problem-solving questions— at least for readers with high, prior knowledge (McNamara et al. 1996). Apparently, a highly coherent text may hinder deeper understanding of high knowledge readers because it reduces their amount of active processing during reading and, as a result, they fail to construct an adequate situation model.

Also when newspaper articles are used updating in episodic memory of situation models is not always effective. In one study, for instance, subjects were presented with a text about the situation in Somalia at the time of the US operation, ‘Restore Hope,’ followed by a second, related text (Van Oostendorp 1996). The first text reported that ‘operation Restore Hope started under American control,’ and in the second text it was said that the command structure of the operation had been changed, such that ‘The United Nations took over the command in Somalia.’ If updating is correct, the old information is replaced by the new information. After reading the second text, readers received an inference test with test items such as ‘The USA troops operate under the UN flag (True False?).’ Results show that the updating performance is, in general, very low. Furthermore, readers who have available an originally appropriate situation model perform a higher degree of updating. That is, the more accurate the original information is represented in memory, the better and faster they judge inferences concerning transformations in the second text. A second, more remarkable, finding is that transformations that are more important to the situation described are less updated than less important transformations. This result has been observed in two experiments with different materials, test items, etc. (Van Oostendorp 1996). It seems that the central part of a situation model may be less easily updated than the peripheral parts. The central part of a mental representation is also less updated when this information is in focus, at least for readers with small initial conceptual networks (Van Oostendorp and Van der Puil 2000). Focus is here manipulated by letting readers compare one text with another along some dimension. The way of examining whether the representation has been updated is based on the cuedassociation task designed by Ferstl and Kintsch (1999). With this task subjects are presented with a word and are asked to provide an association to it. Based on these data, a proximity matrix is calculated for each subject. Subsequently, the similarity of these matrices or conceptual networks, before, as well as after, reading a second text containing new, correcting information, is calculated for the focus group and for a control, nonfocus group. For subjects with a small initial conceptual network there is less updating in the focus group compared to the nonfocus group. These results correspond to what was mentioned previously: changes can be updated less easily with important information, i.e., information in focus, than with less important information.

One important reason why updating may fail is that it is often difficult for readers to discredit old information completely and to exchange that for new information. For example, in reading a story on a fire in a warehouse (Van Oostendorp and Bonebakker 1999), readers in the experimental condition read a sentence such as ‘inflammable materials were carelessly stored in a side room.’ Later, they read that ‘the side room happened to be empty.’ Instead of the sentence ‘inflammable materials were carelessly stored in a side room,’ readers in a control condition received, a neutral sentence, irrelevant to the cause of the fire. The influence of old, obsolete information in the experimental condition could not be fully neutralized by the new, discrediting information. Answers on inference questions, such as what was the cause of the explosion or, for what reason could an insurance company here refuse a claim, were frequently based on the old information, even by subjects who were aware of the fact that information was discredited. Readers in the experimental condition more often gave answers such as ‘because of careless behavior of the owner’ than in the control condition. Recall and direct questions showed that almost all readers had the corrections available but still did not use it during processing of the text. In these experiments, even the explicit instruction that information might be corrected does not lead to a better updating. Readers continue to use the misinformation represented in episodic memory, and keep making inferences based on the corrected information (Johnson and Seifert 1999, Wilkes and Reynolds 1999). It is interesting to know the limits within which readers hold on to old information in memory and don’t take into account new, correcting information. This issue has been explored by Van Oostendorp et al. (in press) using expository text with scientific content. For example, a text that was used, explained a method aimed at increasing the strength of ceramic materials. The old information stated that ‘the treatment (of adding silicon atoms) takes some days,’ while a number of sentences later it was mentioned that ‘by recent advancements the treatment takes as little as a few minutes.’ Inference questions were presented that could tap the interpretation of readers about intermediate events, in order to examine whether the old or the new information source influences this interpretation. For instance, an intermediate sentence contained ‘Reaction speed can be measured by an external device.’ The inference questions were presented after reading the text. The subjects were asked, for example, about the unit of time this device should use to control the process of hardening ceramic materials. Answers of hours or days would mean that readers mainly base their answer on the old information source, as opposed to answers of minutes or seconds, which would mean that they use primarily the new information source. Strengthening of old information in the memory representation—by repeating it in paraphrased form or referring to it indirectly—appears to lead to less updating, that is, to less use of the new information. And, alternatively, readers who read a text in which the new information is reinforced are more likely to use the new information as the basis of inferences. A regulation mechanism seems to be present that is based on weighting evidence favoring either old or new information. According to the outcomes of this evaluation process of the sources, readers choose one or the other point of view, and use that for making inferences, even backwards ones. Thus, in terms of the recent Construction–Integration model of Kintsch (1998), the balance between the strength of sources— the activation values of old and new information in the mental representation—may influence the degree of updating. Subtle reinforcements of old and new information can activate qualitatively different updating strategies, such as holding on to the old information and rejecting the new information, or, on the contrary, switching to a new perspective (as we saw in the studies briefly discussed here).

4. Conclusions

Memory for text depends upon the construction of the mental representation during understanding, more specifically upon the completeness of inferencing and integration, and the extent of updating the mental representation. A number of factors, textual, individual and contextual, are involved with the construction process.

Regarding textual conditions, it appears that texts with semantically highly related concepts led to superficial processing and not noticing errors (as in, e.g., the Moses-illusion experiments and the cat-caught-amouse experiments) and, consequently, to a shallow memory representation. Furthermore, the type of text seems to be important to updating. Expository texts were used in some studies, and readers were able to update their situation models (as in the strengtheningof-ceramic-materials experiments). In contrast, this updating seems to be difficult when stories are used about everyday events (as in the fire-in-warehouse experiments). In addition, the exact character of the correction itself is important. Logical inconsistencies are probably easy to detect but difficult to repair, while a correction or reported change in the world (as in the Somalia experiments) may be more difficult to detect but easy to understand (and to repair). The explicitness, relevance or saliency of old and new information, respectively also influence updating.

Individual characteristics of readers constitute the second block of variables. Relevant here are reading style (Van Oostendorp 1991), prior knowledge, working memory capacity, and beliefs, such as the epistemological belief that integration of ideas implied by a text is important to understanding.

Finally, completeness of processing and updating memory also depends on contextual conditions, such as setting, instruction, and reading goals (Van Oostendorp 1991).

In summary, the processing of readers can often be incomplete in several ways, and these factors have to be taken into account in order to achieve a valid theory of memory for text that can explain the imperfect but also the adaptive memory performance of readers.

Bibliography:

1. Cairns H S, Cowart W, Jablon A D 1981 Effects of prior context upon integration of lexical information during sentence processing. Journal of Verbal Learning and Verbal Behavior 20: 445–53
2. Ferstl E C, Kintsch W 1999 Learning from text: Structural knowledge assessment in the study of discourse comprehension. In: Van Oostendorp H, Goldman S R (eds.) The Construction of Mental Representations During Reading. Erlbaum Associates, Mahwah, NJ, pp. 247–78
3. Hakala C M 1999 Accessibility of spatial information in a situation model. Discourse Processes 27: 261–79
4. Johnson H M, Seifert C M 1999 Modifying mental representations: Comprehending corrections. In: Van Oostendorp H, Goldman S R (eds.) The Construction of Mental Representations During Reading. Erlbaum Associates, Mahwah, NJ, pp. 303–18
5. Kintsch W 1974 The Representation of Meaning in Memory. Erlbaum Associates, Hillsdale, NJ
6. Kintsch W 1998 Comprehension. A Paradigm for Cognition. Cambridge University Press, Cambridge, UK
7. McNamara D S, Kintsch E, Songer N B, Kintsch W 1996 Are good texts always better? Interactions of text coherence, background knowledge, and levels of understanding in learning from text. Cognition and Instruction 14: 1–43
8. Myers J L, O’Brien E J 1998 Accessing the discourse representation during reading. Discourse Processes 26: 131–57
9. Van Oostendorp H 1991 Inferences and integrations made by readers of script-based texts. Journal of Research in Reading 14: 3–21
10. Van Oostendorp H 1994 Text processing in terms of semantic cohesion monitoring. In: Van Oostendorp H, Zwaan R A (eds.) Naturalistic Text Comprehension, Ablex Publishing Corporation, Norwood, NJ, pp. 97–115
11. Van Oostendorp H 1996 Updating situation models derived from newspaper articles. Medienpsychologie. Zeitschrift fur Indi idualund Massenkommunikation 8: 21–33
12. Van Oostendorp H, Bonebakker H 1999 Difficulties in updating mental representations during reading news reports. In: Van Oostendorp H, Goldman S R (eds.) The Construction of Mental Representations During Reading, Erlbaum Associates, Mahwah, NJ, pp. 319–39
13. Van Oostendorp H, De Mul S 1990 Moses beats Adam: A semantic relatedness effect on a semantic illusion. Acta Psychologica 74: 35–46
14. Van Oostendrop H, Otero J, Campanario J in press Conditions of updating during reading. In: Louwerse M, Van Peer W (eds.) Thematics: Interdisciplinary Studies. John Benjamins, Amsterdam
15. Van Oostendorp H, Van der Puil C 2000 The Influence of Focus on Updating a Mental Representation. Proceedings of the 10th Annual Conference of the Society for Text and Discourse, University of Lyon, France
16. Wilkes A L, Reynolds D J 1999 On certain limitations accompanying readers’ interpretations of corrections in episodic text. The Quarterly Journal of Experimental Psychology 52A: 165–83
17. Zwaan R A, Van Oostendorp H 1994 Spatial information and naturalistic story comprehension. In: Van Oostendorp H, Zwaan R A (eds.) Naturalistic Text Comprehension, Ablex Publishing Corporation, Norwood, NJ, pp. 97–115